AI Development & Enterprise Software Intern

I started with a RAG proof-of-concept using Ollama, built so people could query product failure datasets in natural language. I also built a conversational AI agent in Microsoft Copilot Studio that surfaced failure trends across more than 40,000 product-failure records for the Product Support team.

Copilot Studio couldn't reach our organisation's Excel data well at the time, so I requested a Dataverse connector and tried to expose it to the agent through MCP. It had its own ceiling of 5,000 rows per query, but it still worked better than attaching the Excel file directly as a source. The other half of the problem was the agent itself. I had to keep it strictly to its system prompt and get it to understand what the rows and columns meant, which ended up meaning labelling columns by hand with context so it could read the data properly.

The inventory app started with watching my team struggle to track the hardware products they loaned out. Everything lived in an Excel sheet that was poorly filled in and barely documented, with missing parts and unclear ownership. I took this on myself, looked at the pain points, and built a PHP and Docker app with the web development, containerization, and networking I already knew. I set it up on a static local IP so anyone on the office wifi could reach it. It tracks owner, product type, and product class, supports loan approvals, and sends automatic reminders based on loan due dates through cron jobs. It ended up managing more than 400 products across 8 categories for 25-plus staff.

I also automated JIRA ticket validation with regex-based logic across the 30 to 50 tickets that came in daily. The automated content checks cut escalation-processing time by more than 15 minutes per ticket, which adds up quickly at that volume.